Enhancing Domain-Specific ASR via Dual-Stage RAG-Based Context Injection
Abstract
Standard Automatic Speech Recognition (ASR) systems struggle to accurately transcribe dense, domain-specific terminology in academic lectures. This paper introduces a dual-stage context injection pipeline that enhances transcription quality through a multi-stage Retrieval-Augmented Generation (RAG) framework. Domain knowledge (e.g., lecture slides) is incorporated both during decoding via prompt-based biasing and during post-processing through Large Language Model (LLM)based refinement. Through a factorial evaluation grid yielding $N=27$ experimental runs across three academic domains, we reveal a systematic trade-off between verbatim accuracy and semantic utility. While biasing reduces the Word Error Rate (WER) as low as 1.5% in best-case configurations (EELarge-v3 Stage 1), the subsequent refinement stage increases WER due to structural transformations aimed at reducing extraneous cognitive load. We characterize this phenomenon as an observed refinement paradox, where a higher WER-driven by linguistic normalization-coincides with improved semantic fidelity and instructional quality. To measure this effect, we utilize the Knowledge Recovery Rate (KRR), achieving values up to 0.96 in preserving domainspecific k n owledge. Furthermore, our results in dicate a Contextual Compensation Effect, in which context-aware refinement enables smaller ASR models to match or exceed the semantic performance of larger backbones in pedagogical summarization. This suggests that contextual grounding can be a decisive factor, potentially outweighing model size in specialized, resourceconstrained educational environments. The proposed system is implemented as a privacy-compliant, on-premise solution.